Fix Whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116173 91177308-0d34-0410-b5e6-96231b3b80d8

force a triple, varargs isn't supported with the SVR4 ABI the buildbot tells me.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116170 91177308-0d34-0410-b5e6-96231b3b80d8

fix the default va_arg expansion (in the realignment case) to not implicitly
truncate the stack pointer to 32-bits on a 64-bit machine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116169 91177308-0d34-0410-b5e6-96231b3b80d8

fix the expansion of va_arg instruction on PPC to know the arg
alignment for PPC32/64, avoiding some masking operations.

llvm-gcc expands vaarg inline instead of using the instruction
so it has never hit this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116168 91177308-0d34-0410-b5e6-96231b3b80d8

clarify that zero sized vectors are illegal, PR8340

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116167 91177308-0d34-0410-b5e6-96231b3b80d8

Now using a variant of the existing inlining heuristics to decide whether to create a given specialization of a function in PartialSpecialization.  If the total performance bonus across all callsites passing the same constant exceeds the specialization cost, we create the specialization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116158 91177308-0d34-0410-b5e6-96231b3b80d8

Fix dead link.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116157 91177308-0d34-0410-b5e6-96231b3b80d8

Silence compiler warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116156 91177308-0d34-0410-b5e6-96231b3b80d8

MC-COFF: Fix .bss section size. Fixes PR8335. Patch by NAKAMUTA Takumi!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116155 91177308-0d34-0410-b5e6-96231b3b80d8

Don't test a removed function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116154 91177308-0d34-0410-b5e6-96231b3b80d8

syntax-highlighting: Fix module asm keyword.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116152 91177308-0d34-0410-b5e6-96231b3b80d8

MC-COFF: Implement InitSections. Fixes PR8335.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116151 91177308-0d34-0410-b5e6-96231b3b80d8

MC-COFF: Add COFFAsmParser. Completes PR8343.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116150 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116149 91177308-0d34-0410-b5e6-96231b3b80d8

MC-COFF: Assert on non-coff sections.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116148 91177308-0d34-0410-b5e6-96231b3b80d8

Add Kate syntax highlighting files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116146 91177308-0d34-0410-b5e6-96231b3b80d8

Add VLD4 scheduling itineraries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116143 91177308-0d34-0410-b5e6-96231b3b80d8

MC: Move ParseDirectiveELFType into ELFAsmParser. COFF uses .type for something else.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116142 91177308-0d34-0410-b5e6-96231b3b80d8

Finish vld3 and vld4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116140 91177308-0d34-0410-b5e6-96231b3b80d8

Complete vld2 instruction itineries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116136 91177308-0d34-0410-b5e6-96231b3b80d8

Multiply instructions are issued on pipeline 0. They do not need to reserve pipeline 1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116135 91177308-0d34-0410-b5e6-96231b3b80d8

Correct some load / store instruction itinerary mistakes:
1. Cortex-A8 load / store multiplies can only issue on ALU0.
2. Eliminate A8_Issue, A8_LSPipe will correctly limit the load / store issues.
3. Correctly model all vld1 and vld2 variants.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116134 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify test and move into a generic "crash" ll file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116130 91177308-0d34-0410-b5e6-96231b3b80d8

Check to make sure that the iterator isn't at the beginning of the basic block
before decrementing. <rdar://problem/8529919>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116126 91177308-0d34-0410-b5e6-96231b3b80d8

add jit support for the new psuedo instructions I added for
the add/or xform.  The JIT isn't mcized yet, boo.

This fixes Olden/voronoi, bh and a ton of other stuff that
uses the jit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116125 91177308-0d34-0410-b5e6-96231b3b80d8

machine a mutable machineinstr down into emitInstruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116124 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the store part of this as well.  Fixes smg2000.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116123 91177308-0d34-0410-b5e6-96231b3b80d8

Rename SplitEditor::rewrite to finish() and break it out into a couple of new
functions: computeRemainder and rewrite.

When the remainder breaks up into multiple components, remember to rewrite those
uses as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116121 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid compiler warning: comparison between signed and unsigned integer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116119 91177308-0d34-0410-b5e6-96231b3b80d8

Extract method ProcessUses from LocalRewriter::RewriteMBB. Both parent and child
are still way too long, but it's a start.

No functional change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116116 91177308-0d34-0410-b5e6-96231b3b80d8

Do not check that the bodies of two defs of same linkonce global are the same.
Such a check does not make any sense in presense of inlining and other compiler-dependent stuff.
This should fix bunch of warnings on mingw32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116113 91177308-0d34-0410-b5e6-96231b3b80d8

Implement a few more binary encoding bits. Still very early stage proof-of-
concept level stuff at this point, but it is generally working for those
instructions that know how to map the operands.

This patch fills in the register operands for add/sub/or/etc instructions
and adds the conditional execution predicate encoding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116112 91177308-0d34-0410-b5e6-96231b3b80d8

Allow << streaming of MCOperand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116107 91177308-0d34-0410-b5e6-96231b3b80d8

Classify value numbers into connected components in linear time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116105 91177308-0d34-0410-b5e6-96231b3b80d8

Call InitSections in llc and clang so that the binaries produced by them
are easier to diff with those produced by llvm-mc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116095 91177308-0d34-0410-b5e6-96231b3b80d8

Filter out illegal formulae after updating offsets, not before, so that
formulae which become illegal as a result of the offset updating don't
escape.

This is for rdar://8529692. No testcase yet, because the given cases
hit use-list ordering differences.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116093 91177308-0d34-0410-b5e6-96231b3b80d8

Recommit 116056, now with the missing file...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116083 91177308-0d34-0410-b5e6-96231b3b80d8

Don't waste time unfolding simple loads. The unfolded copy won't be hoisted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116081 91177308-0d34-0410-b5e6-96231b3b80d8

Fix operand latency computation in cases where the definition operand is
implicit. e.g.
%D6<def>, %D7<def> = VLD1q16 %R2<kill>, 0, ..., %Q3<imp-def>
%Q1<def> = VMULv8i16 %Q1<kill>, %Q3<kill>, ...

The real definition indices are 0,1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116080 91177308-0d34-0410-b5e6-96231b3b80d8

CrashRecovery: Fix raise() override to actually send the right signal, *cough*.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116072 91177308-0d34-0410-b5e6-96231b3b80d8

A few 80 column cleanups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116069 91177308-0d34-0410-b5e6-96231b3b80d8

trailing whitespace

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116068 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply 116059, this time without the fatfingered pasto at the top.

''const'ify getMachineOpValue() and associated helpers.'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116067 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting 116059. Bots are unhappy with it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116064 91177308-0d34-0410-b5e6-96231b3b80d8

reverting 116056: win64_params.ll may need to be conditionalized?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116063 91177308-0d34-0410-b5e6-96231b3b80d8

Line number 0 indicates there is no source line/file name info available for this construct.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116061 91177308-0d34-0410-b5e6-96231b3b80d8

'const'ify getMachineOpValue() and associated helpers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116059 91177308-0d34-0410-b5e6-96231b3b80d8

Start separating out code metrics into code size metrics and code performance metrics.  Partial Specialization will apply the former to function specializations, and the latter to all callsites that can use a specialization, in order to decide whether to create a specialization

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116057 91177308-0d34-0410-b5e6-96231b3b80d8

Small patch to restore home register stack space allocation for the Win64 case.  Add test case.  This code eventually needs to be tighter, since it's always allocating it, even in leaf routines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116056 91177308-0d34-0410-b5e6-96231b3b80d8

Change register allocation order for ARM VFP and NEON registers to put the
callee-saved registers at the end of the lists.  Also prefer to avoid using
the low registers that are in register subclasses required by certain
instructions, so that those registers will more likely be available when needed.
This change makes a huge improvement in spilling in some cases.  Thanks to
Jakob for helping me realize the problem.

Most of this patch is fixing the testsuite.  There are quite a few places
where we're checking for specific registers.  I changed those to wildcards
in places where that doesn't weaken the tests.  The spill-q.ll and
thumb2-spill-q.ll tests stopped spilling with this change, so I added a bunch
of live values to force spills on those tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116055 91177308-0d34-0410-b5e6-96231b3b80d8

testcase that goes with r116053

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116054 91177308-0d34-0410-b5e6-96231b3b80d8

fix a subtle bug I introduced in my refactoring, where we stopped preferring
the i8 versions of instructions in some cases.  In test6, we started
generating:

cmpq $0, -8(%rsp)            ## encoding: [0x48,0x81,0x7c,0x24,0xf8,0x00,0x00,0x00,0x00]
                                        ## <MCInst #478 CMP64mi32
                                        ##  <MCOperand Reg:114>
                                        ##  <MCOperand Imm:1>
                                        ##  <MCOperand Reg:0>
                                        ##  <MCOperand Imm:-8>
                                        ##  <MCOperand Reg:0>
                                        ##  <MCOperand Imm:0>>

instead of:

cmpq $0, -8(%rsp)            ## encoding: [0x48,0x83,0x7c,0x24,0xf8,0x00]
                                        ## <MCInst #479 CMP64mi8
                                        ##  <MCOperand Reg:114>
                                        ##  <MCOperand Imm:1>
                                        ##  <MCOperand Reg:0>
                                        ##  <MCOperand Imm:-8>
                                        ##  <MCOperand Reg:0>
                                        ##  <MCOperand Imm:0>>

Fix this and add some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116053 91177308-0d34-0410-b5e6-96231b3b80d8

rename test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116052 91177308-0d34-0410-b5e6-96231b3b80d8

merge tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116051 91177308-0d34-0410-b5e6-96231b3b80d8

filecheckize.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116050 91177308-0d34-0410-b5e6-96231b3b80d8

reapply: Use the new TB_NOT_REVERSABLE flag instead of special
reapply: reimplement the second half of the or/add optimization.  We should now

with no changes.  Turns out that one missing "Defs = [EFLAGS]" can upset things
a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116040 91177308-0d34-0410-b5e6-96231b3b80d8

reapply the patch reverted in r116033:
"Reimplement (part of) the or -> add optimization.  Matching 'or' into 'add'"

With a critical fix: the add pseudos clobber EFLAGS.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116039 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Formatting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116038 91177308-0d34-0410-b5e6-96231b3b80d8

MC-COFF: Fix Simple and Complex type. Fixes PR8320.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116037 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116036 91177308-0d34-0410-b5e6-96231b3b80d8

Update CMake.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116034 91177308-0d34-0410-b5e6-96231b3b80d8

Revert "Reimplement (part of) the or -> add optimization.  Matching 'or' into
'add'", which seems to have broken just about everything.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116033 91177308-0d34-0410-b5e6-96231b3b80d8

Revert "Use the new TB_NOT_REVERSABLE flag instead of special ", which depends
on r116007, which I am about to revert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116032 91177308-0d34-0410-b5e6-96231b3b80d8

Revert "reimplement the second half of the or/add optimization.  We should now",
which depends on r116007, which I am about to revert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116031 91177308-0d34-0410-b5e6-96231b3b80d8

Fix -Asserts warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116030 91177308-0d34-0410-b5e6-96231b3b80d8

Move to thumb2 loads, fixes a problem with incoming registers
as thumb1.

Fixes lencod.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116027 91177308-0d34-0410-b5e6-96231b3b80d8

reimplement the second half of the or/add optimization.  We should now
only end up emitting LEA instead of OR.  If we aren't able to promote
something into an LEA, we should never be emitting it as an ADD.

Add some testcases that we emit "or" in cases where we used to produce
an "add".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116026 91177308-0d34-0410-b5e6-96231b3b80d8

Add test file for simple ARM binary encodings with MC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116024 91177308-0d34-0410-b5e6-96231b3b80d8

Enable binary encoding of some simple instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116022 91177308-0d34-0410-b5e6-96231b3b80d8

Try to fix ocaml bindings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116021 91177308-0d34-0410-b5e6-96231b3b80d8

Make <target>CodeEmitter::getBinaryCodeForInstr() a const method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116018 91177308-0d34-0410-b5e6-96231b3b80d8

Use the new TB_NOT_REVERSABLE flag instead of special
casing FsMOVAPDrr/FsMOVAPSrr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116016 91177308-0d34-0410-b5e6-96231b3b80d8

MC-COFF: Add test for my last commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116015 91177308-0d34-0410-b5e6-96231b3b80d8

simplify some map operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116014 91177308-0d34-0410-b5e6-96231b3b80d8

MC-COFF: Handle relaxation in COFF better. Fixes PR8321.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116013 91177308-0d34-0410-b5e6-96231b3b80d8

Delete the FormulaSorter class and inline its one method into its
one user. This code will be restructured soon and FormulaSorter
is getting in the way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116012 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a spello.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116011 91177308-0d34-0410-b5e6-96231b3b80d8

Charge a formula for explicit multiplies on scaled registers too,
not just base registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116010 91177308-0d34-0410-b5e6-96231b3b80d8

Use size_t for consistency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116009 91177308-0d34-0410-b5e6-96231b3b80d8

When merging one use into another, transfer the offsets from
the old use to the new one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116008 91177308-0d34-0410-b5e6-96231b3b80d8

Reimplement (part of) the or -> add optimization.  Matching 'or' into 'add'
is general goodness because it allows ORs to be converted to LEA to avoid
inserting copies.  However, this is bad because it makes the generated .s
file less obvious and gives valgrind heartburn (tons of false positives in
bitfield code).

While the general fix should be in valgrind, we can at least try to avoid
emitting ADD instructions that *don't* get promoted to LEA.  This is more
work because it requires introducing pseudo instructions to represents
"add that knows the bits are disjoint", but hey, people really love valgrind.

This fixes this testcase:
https://bugs.kde.org/show_bug.cgi?id=242137#c20

the add r/i cases are coming next.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116007 91177308-0d34-0410-b5e6-96231b3b80d8

After splitting, the remaining LiveInterval may be fragmented into multiple
connected components. These components should be allocated different virtual
registers because there is no reason for them to be allocated together.

Add the ConnectedVNInfoEqClasses class to calculate the connected components,
and move values to new LiveIntervals.

Use it from SplitKit::rewrite by creating new virtual registers for the
components.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116006 91177308-0d34-0410-b5e6-96231b3b80d8

Fix LSR to keep the RegUseTracker up to date when combining users.
This doesn't usually matter, because the other heuristics usually
succeed regardless, but it's good to keep the register use
bookkeeping consistent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116005 91177308-0d34-0410-b5e6-96231b3b80d8

Remove LoopIndexSplit pass. It is neither maintained nor used by anyone.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116004 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak cmake build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116003 91177308-0d34-0410-b5e6-96231b3b80d8

Code refactoring.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116002 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce casting in various tables by defining the table
with the right types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116001 91177308-0d34-0410-b5e6-96231b3b80d8

Warn about broken GCCs on ARM due to the problem fixed:
http://gcc.gnu.org/ml/gcc-patches/2010-09/msg01070.html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115998 91177308-0d34-0410-b5e6-96231b3b80d8

simplify code: don't build up vector only to assert it is empty.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115997 91177308-0d34-0410-b5e6-96231b3b80d8

Now with fewer extraneous semicolons!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115996 91177308-0d34-0410-b5e6-96231b3b80d8

Move checking for t2MOVCCi16 to the right place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115994 91177308-0d34-0410-b5e6-96231b3b80d8

Trivial MC code emitter shell. No instruction forms actually handled yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115993 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly check if a path is a directory. Fix by Brian Korver.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115991 91177308-0d34-0410-b5e6-96231b3b80d8

Provie a clearner interface so that FE can decide whether a function has prototype or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115988 91177308-0d34-0410-b5e6-96231b3b80d8

Include the auto-generated bits for machine encoding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115987 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo in comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115986 91177308-0d34-0410-b5e6-96231b3b80d8

Remember to promote load/store types for stack to register size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115984 91177308-0d34-0410-b5e6-96231b3b80d8

convert test to use the existing classes that the multipatterns
use.  Since TEST is completely different than all other binops,
don't define a multipattern for it.

This completes factorization of binops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115982 91177308-0d34-0410-b5e6-96231b3b80d8

convert cmp to use a multipattern

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115978 91177308-0d34-0410-b5e6-96231b3b80d8

Canonicalize X86ISD::MOVDDUP nodes to v2f64 to make sure all cases match. Also eliminate unneeded isel patterns. rdar://8520311

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115977 91177308-0d34-0410-b5e6-96231b3b80d8